The AI honeymoon is over. After years of free trials, flat-fee subscriptions, and "unlimited" usage, the bill has arrived — and corporate executives are in shock.
A new survey from KPMG, the global consultancy giant, reveals that 29% of senior executives across 20 countries are struggling to understand and control the operating costs of their AI deployments. The survey of more than 2,000 C-suite leaders found that nearly half are actively re-phasing — slowing down or rethinking — their AI strategies when the costs start to outweigh the expected value.
From All-You-Can-Eat to Pay-Per-Token
The root cause is a fundamental shift in how AI vendors charge. Anthropic, OpenAI, and GitHub have all moved away from flat-fee subscription models to usage-based billing based on tokens. The strategy was classic drug-dealer economics: get users hooked with free or cheap access, then raise prices once they're dependent.
It worked. Engineers across enterprises have integrated AI coding assistants, chatbots, and agents into their daily workflows. They've stopped memorizing syntax, stopped writing boilerplate code, and in many cases, stopped thinking through problems the way they used to. The AI companies got the user lock-in they wanted. Now they're monetizing it.
The problem? Nobody bothered to calculate what "dependent on AI" would actually cost at scale.
The Numbers Are Ugly
Gartner research published in June 2026 delivered a sobering forecast: AI coding agents could soon cost more than the developers using them. By 2028, the average global cost of AI coding assistance per developer is projected to exceed the average global developer salary.
This isn't a future problem — it's already happening in lower-wage markets. In India, where developer salaries are significantly below the global average, the cost of AI agents has already surpassed what companies pay their human engineers. The math is brutal: AI agent costs are the same worldwide, but human salaries vary dramatically by geography.
And there's no standardization. Each model provider bills differently. Some charge per token, some per request, some per feature. Comparing costs across vendors is nearly impossible, and the lack of transparency means finance teams can't forecast, procurement can't negotiate, and CTOs can't budget.
The Productivity Paradox
Here's the kicker: there's no direct relationship between token consumption and productivity gains. Gartner found that companies using the most AI tokens weren't necessarily shipping more code, closing more tickets, or generating more revenue. In fact, the research suggested that controlled, careful AI usage produced better code quality than unrestrained "tokenmaxxing."
Spencer Kimball, CEO of Cockroach Labs and a veteran of Google's engineering culture, put it bluntly: "There's no point in maxing out a model when you haven't provided the right context because you'll just get more rubbish back." His company deliberately avoids tokenmaxxing, mixing cheap open-source models with commercial ones and only using expensive AI when the context justifies it.
The enterprise AI market is discovering what the cloud market learned a decade ago: unlimited usage sounds great until you see the bill. And unlike cloud infrastructure, where costs generally correlate with business growth, AI costs can spiral without delivering proportional value.
The Open Source Fightback
The backlash is creating opportunities for open-source tools and efficiency startups. One standout is Project Headroom, created by a Netflix engineer and open-sourced in May 2026. The tool trims redundant input tokens before they're sent to LLMs — removing unnecessary formatting, stripping repetitive context, and compressing prompts.
The savings are dramatic. Users report cutting 90% of server logs from prompts, 70% of JSON formatting overhead, and overall token reductions of 30-50%. For one enterprise user, the tool saved $700,000 in annual AI costs. The philosophy is simple: "tokenminning" instead of "tokenmaxxing."
Database vendors are also jumping in. Pinecone, the vector database company, is building semantic layers that sit between AI agents and business data. Instead of calling an LLM every time an agent needs to understand a database schema or financial process, the semantic layer stores that knowledge locally. Fewer LLM calls, lower costs, faster responses.
The Macro Picture: $1.5 Trillion at Stake
The cost crunch isn't just a vendor pricing problem — it's an infrastructure problem. One major investment house estimates that $1.5 trillion will be spent on AI data centers between 2025 and 2030. That money has to come from somewhere, and right now, it's coming from enterprise AI budgets that are proving smaller than expected.
Oracle is perhaps the most exposed. The company announced $450 billion in committed data center spending last September, with OpenAI reportedly on the hook for $300 billion of that. Oracle is borrowing heavily to build this infrastructure. If enterprise AI spending slows — which the KPMG and Gartner data suggest it will — Oracle could be left with massive capacity and insufficient demand.
S&P Global has already downgraded Oracle to BBB-, one notch above junk status, citing OpenAI as a "key credit risk." The hyperscalers — Amazon, Google, Microsoft — have more diversified businesses to absorb a potential AI spending slowdown. Oracle doesn't.
🔥 Hot Takes
1. The AI industry's pricing model is fundamentally broken because it was designed to create addiction, not value. Vendors gave away AI for free or cheap to build dependency, then switched to metered billing once users couldn't live without it. This is the same playbook as cigarette companies, mobile games, and SaaS freemium models. The difference is that AI costs scale unpredictably — a developer's "helpful assistant" can become a six-figure annual expense without anyone noticing until the invoice arrives.
2. The "tokenmaxxing" culture proves that AI efficiency research has been dangerously neglected. For all the billions poured into making models bigger and more capable, almost nobody has been working on making them cheaper to use. The AI labs optimized for benchmarks, not budgets. Now that enterprises are hitting cost walls, the industry is scrambling to catch up on efficiency — but the inefficiency is already baked into the models, the APIs, and the user habits.
3. This is the beginning of AI's "cloud cost optimization" era — and it will reshape the industry as much as cloud optimization reshaped AWS. Remember when companies were shocked by their first AWS bills? It created an entire industry of cloud cost management tools, FinOps practices, and multi-cloud strategies. AI is about to go through the same transformation. The winners won't be the companies with the biggest models — they'll be the companies that help enterprises use AI efficiently. The "tokenminning" revolution is coming, and it will be bigger than the "tokenmaxxing" boom that preceded it.
Bottom line: The AI industry sold enterprises a dream of unlimited productivity at a fixed cost. The reality is metered billing, unpredictable costs, and diminishing returns. The companies that survive this transition will be the ones that stop treating AI like a magic wand and start treating it like what it is: an expensive, powerful tool that requires careful management, clear ROI measurement, and ruthless cost optimization. The free lunch is over. Now comes the hard part.